Style Customization of Text-to-Vector Generation with Image Diffusion Priors

Setup

Create a new conda environment:

conda create --name svg_diffusion python=3.10
conda activate svg_diffusion

Install pytorch and the following libraries:

conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch
pip install -r requirements.txt

Install diffvg:

git clone https://github.com/BachiLi/diffvg.git
cd diffvg
git submodule update --init --recursive
conda install -y -c anaconda cmake
conda install -y -c conda-forge ffmpeg
pip install svgwrite svgpathtools cssutils torch-tools
python setup.py install

TODO

• Release Stage1 model weights.
• Release Stage2 model weights.
• Release the inference code.
• Release training scripts.

Model Weights

Model name	Weight
Stage1	link
Stage2	TBA

Quickstart

1. Download the pretrained weights
   Put the files in the pretrained directory.

2. Launch the Gradio demo:

   CUDA_VISIBLE_DEVICES=0 python -m svg_ldm.gradio_t2svg
   

   Use the Generation Count slider to sample multiple SVGs in one click.

3. or run the inference script:

   CUDA_VISIBLE_DEVICES=0 python -m svg_ldm.test_ddpm_tr_svgs
   

   Adjust test_num to control how many samples are produced per prompt.

Tips

• Diffusion sampling is stochastic. Vary the random seed or adjust the sampling settings to explore different outputs.
• We recommend DDPM sampling for higher-quality results, which takes about 30 seconds to generate one SVG on an NVIDIA RTX 4090.
• The model was trained only on simple class labels (see dataset/label.csv), so it doesn’t understand complex text prompts. Fine-tune the model on richer SVG datasets can support more detailed prompts.

Contact

If you have any question, contact us through email at zhangpeiying17@gmail.com.